Merge "Renaming vp9_short_idct10_16x16 to vp9_short_idct16x16_10."
diff --git a/build/make/configure.sh b/build/make/configure.sh
index bb7ab41..f361021 100755
--- a/build/make/configure.sh
+++ b/build/make/configure.sh
@@ -1062,7 +1062,7 @@
setup_gnu_toolchain
add_cflags -use-msasm -use-asm
add_ldflags -i-static
- enabled x86_64 && add_cflags -ipo -no-prec-div -static -xSSE2 -axSSE2
+ enabled x86_64 && add_cflags -ipo -static -O3
enabled x86_64 && AR=xiar
case ${tune_cpu} in
atom*)
diff --git a/test/intrapred_test.cc b/test/intrapred_test.cc
index f5f6d5b..5fba700 100644
--- a/test/intrapred_test.cc
+++ b/test/intrapred_test.cc
@@ -34,13 +34,17 @@
}
protected:
- void SetupMacroblock(uint8_t *data, int block_size, int stride,
+ void SetupMacroblock(MACROBLOCKD *mbptr,
+ MODE_INFO *miptr,
+ uint8_t *data,
+ int block_size,
+ int stride,
int num_planes) {
- memset(&mb_, 0, sizeof(mb_));
- memset(&mi_, 0, sizeof(mi_));
- mb_.up_available = 1;
- mb_.left_available = 1;
- mb_.mode_info_context = &mi_;
+ mbptr_ = mbptr;
+ miptr_ = miptr;
+ mbptr_->up_available = 1;
+ mbptr_->left_available = 1;
+ mbptr_->mode_info_context = miptr_;
stride_ = stride;
block_size_ = block_size;
num_planes_ = num_planes;
@@ -63,14 +67,14 @@
virtual void Predict(MB_PREDICTION_MODE mode) = 0;
void SetLeftUnavailable() {
- mb_.left_available = 0;
+ mbptr_->left_available = 0;
for (int p = 0; p < num_planes_; p++)
for (int i = -1; i < block_size_; ++i)
data_ptr_[p][stride_ * i - 1] = 129;
}
void SetTopUnavailable() {
- mb_.up_available = 0;
+ mbptr_->up_available = 0;
for (int p = 0; p < num_planes_; p++)
memset(&data_ptr_[p][-1 - stride_], 127, block_size_ + 2);
}
@@ -96,13 +100,13 @@
for (int p = 0; p < num_planes_; p++) {
// calculate expected DC
int expected;
- if (mb_.up_available || mb_.left_available) {
- int sum = 0, shift = BlockSizeLog2Min1() + mb_.up_available +
- mb_.left_available;
- if (mb_.up_available)
+ if (mbptr_->up_available || mbptr_->left_available) {
+ int sum = 0, shift = BlockSizeLog2Min1() + mbptr_->up_available +
+ mbptr_->left_available;
+ if (mbptr_->up_available)
for (int x = 0; x < block_size_; x++)
sum += data_ptr_[p][x - stride_];
- if (mb_.left_available)
+ if (mbptr_->left_available)
for (int y = 0; y < block_size_; y++)
sum += data_ptr_[p][y * stride_ - 1];
expected = (sum + (1 << (shift - 1))) >> shift;
@@ -209,8 +213,8 @@
}
}
- MACROBLOCKD mb_;
- MODE_INFO mi_;
+ MACROBLOCKD *mbptr_;
+ MODE_INFO *miptr_;
uint8_t *data_ptr_[2]; // in the case of Y, only [0] is used
int stride_;
int block_size_;
@@ -228,12 +232,18 @@
protected IntraPredBase {
public:
static void SetUpTestCase() {
+ mb_ = reinterpret_cast<MACROBLOCKD*>(
+ vpx_memalign(32, sizeof(MACROBLOCKD)));
+ mi_ = reinterpret_cast<MODE_INFO*>(
+ vpx_memalign(32, sizeof(MODE_INFO)));
data_array_ = reinterpret_cast<uint8_t*>(
vpx_memalign(kDataAlignment, kDataBufferSize));
}
static void TearDownTestCase() {
vpx_free(data_array_);
+ vpx_free(mi_);
+ vpx_free(mb_);
data_array_ = NULL;
}
@@ -250,12 +260,12 @@
virtual void SetUp() {
pred_fn_ = GetParam();
- SetupMacroblock(data_array_, kBlockSize, kStride, 1);
+ SetupMacroblock(mb_, mi_, data_array_, kBlockSize, kStride, 1);
}
virtual void Predict(MB_PREDICTION_MODE mode) {
- mb_.mode_info_context->mbmi.mode = mode;
- REGISTER_STATE_CHECK(pred_fn_(&mb_,
+ mbptr_->mode_info_context->mbmi.mode = mode;
+ REGISTER_STATE_CHECK(pred_fn_(mbptr_,
data_ptr_[0] - kStride,
data_ptr_[0] - 1, kStride,
data_ptr_[0], kStride));
@@ -263,8 +273,12 @@
intra_pred_y_fn_t pred_fn_;
static uint8_t* data_array_;
+ static MACROBLOCKD * mb_;
+ static MODE_INFO *mi_;
};
+MACROBLOCKD* IntraPredYTest::mb_ = NULL;
+MODE_INFO* IntraPredYTest::mi_ = NULL;
uint8_t* IntraPredYTest::data_array_ = NULL;
TEST_P(IntraPredYTest, IntraPredTests) {
@@ -299,12 +313,18 @@
protected IntraPredBase {
public:
static void SetUpTestCase() {
+ mb_ = reinterpret_cast<MACROBLOCKD*>(
+ vpx_memalign(32, sizeof(MACROBLOCKD)));
+ mi_ = reinterpret_cast<MODE_INFO*>(
+ vpx_memalign(32, sizeof(MODE_INFO)));
data_array_ = reinterpret_cast<uint8_t*>(
vpx_memalign(kDataAlignment, kDataBufferSize));
}
static void TearDownTestCase() {
vpx_free(data_array_);
+ vpx_free(mi_);
+ vpx_free(mb_);
data_array_ = NULL;
}
@@ -322,12 +342,12 @@
virtual void SetUp() {
pred_fn_ = GetParam();
- SetupMacroblock(data_array_, kBlockSize, kStride, 2);
+ SetupMacroblock(mb_, mi_, data_array_, kBlockSize, kStride, 2);
}
virtual void Predict(MB_PREDICTION_MODE mode) {
- mb_.mode_info_context->mbmi.uv_mode = mode;
- pred_fn_(&mb_, data_ptr_[0] - kStride, data_ptr_[1] - kStride,
+ mbptr_->mode_info_context->mbmi.uv_mode = mode;
+ pred_fn_(mbptr_, data_ptr_[0] - kStride, data_ptr_[1] - kStride,
data_ptr_[0] - 1, data_ptr_[1] - 1, kStride,
data_ptr_[0], data_ptr_[1], kStride);
}
@@ -340,8 +360,12 @@
// We use 9 lines so we have one line above us for top-prediction.
// [0] = U, [1] = V
static uint8_t* data_array_;
+ static MACROBLOCKD* mb_;
+ static MODE_INFO* mi_;
};
+MACROBLOCKD* IntraPredUVTest::mb_ = NULL;
+MODE_INFO* IntraPredUVTest::mi_ = NULL;
uint8_t* IntraPredUVTest::data_array_ = NULL;
TEST_P(IntraPredUVTest, IntraPredTests) {
diff --git a/vp8/common/x86/vp8_asm_stubs.c b/vp8/common/x86/vp8_asm_stubs.c
index c0416b7..b409293 100644
--- a/vp8/common/x86/vp8_asm_stubs.c
+++ b/vp8/common/x86/vp8_asm_stubs.c
@@ -611,16 +611,12 @@
for (r = 0; r < 4; r++)
{
- #if !(CONFIG_FAST_UNALIGNED)
dst_ptr[0] = src_ptr[0];
dst_ptr[1] = src_ptr[1];
dst_ptr[2] = src_ptr[2];
dst_ptr[3] = src_ptr[3];
- #else
- *(uint32_t *)dst_ptr = *(uint32_t *)src_ptr ;
- #endif
- dst_ptr += dst_pitch;
- src_ptr += src_pixels_per_line;
+ dst_ptr += dst_pitch;
+ src_ptr += src_pixels_per_line;
}
}
}
diff --git a/vp8/encoder/dct.c b/vp8/encoder/dct.c
index b5a11ae..091554a 100644
--- a/vp8/encoder/dct.c
+++ b/vp8/encoder/dct.c
@@ -20,10 +20,10 @@
for (i = 0; i < 4; i++)
{
- a1 = ((ip[0] + ip[3])<<3);
- b1 = ((ip[1] + ip[2])<<3);
- c1 = ((ip[1] - ip[2])<<3);
- d1 = ((ip[0] - ip[3])<<3);
+ a1 = ((ip[0] + ip[3]) * 8);
+ b1 = ((ip[1] + ip[2]) * 8);
+ c1 = ((ip[1] - ip[2]) * 8);
+ d1 = ((ip[0] - ip[3]) * 8);
op[0] = a1 + b1;
op[2] = a1 - b1;
@@ -72,10 +72,10 @@
for (i = 0; i < 4; i++)
{
- a1 = ((ip[0] + ip[2])<<2);
- d1 = ((ip[1] + ip[3])<<2);
- c1 = ((ip[1] - ip[3])<<2);
- b1 = ((ip[0] - ip[2])<<2);
+ a1 = ((ip[0] + ip[2]) * 4);
+ d1 = ((ip[1] + ip[3]) * 4);
+ c1 = ((ip[1] - ip[3]) * 4);
+ b1 = ((ip[0] - ip[2]) * 4);
op[0] = a1 + d1 + (a1!=0);
op[1] = b1 + c1;
diff --git a/vp9/common/arm/neon/vp9_idct32x32_neon.c b/vp9/common/arm/neon/vp9_idct32x32_neon.c
deleted file mode 100644
index ceecd6f..0000000
--- a/vp9/common/arm/neon/vp9_idct32x32_neon.c
+++ /dev/null
@@ -1,47 +0,0 @@
-/*
- * Copyright (c) 2013 The WebM project authors. All Rights Reserved.
- *
- * Use of this source code is governed by a BSD-style license
- * that can be found in the LICENSE file in the root of the source
- * tree. An additional intellectual property rights grant can be found
- * in the file PATENTS. All contributing project authors may
- * be found in the AUTHORS file in the root of the source tree.
- */
-
-#include "vp9/common/vp9_common.h"
-
-// defined in vp9/common/arm/neon/vp9_short_idct32x32_add_neon.asm
-extern void idct32_transpose_and_transform(int16_t *transpose_buffer,
- int16_t *output, int16_t *input);
-extern void idct32_combine_add(uint8_t *dest, int16_t *out, int dest_stride);
-
-
-// defined in vp9/common/arm/neon/vp9_short_idct16x16_add_neon.asm
-extern void save_neon_registers();
-extern void restore_neon_registers();
-
-void vp9_short_idct32x32_add_neon(int16_t *input, uint8_t *dest,
- int dest_stride) {
- // TODO(cd): move the creation of these buffers within the ASM file
- // internal buffer used to transpose 8 lines into before transforming them
- int16_t transpose_buffer[32 * 8];
- // results of the first pass (transpose and transform rows)
- int16_t pass1[32 * 32];
- // results of the second pass (transpose and transform columns)
- int16_t pass2[32 * 32];
-
- // save register we need to preserve
- save_neon_registers();
- // process rows
- idct32_transpose_and_transform(transpose_buffer, pass1, input);
- // process columns
- // TODO(cd): do these two steps/passes within the ASM file
- idct32_transpose_and_transform(transpose_buffer, pass2, pass1);
- // combine and add to dest
- // TODO(cd): integrate this within the last storage step of the second pass
- idct32_combine_add(dest, pass2, dest_stride);
- // restore register we need to preserve
- restore_neon_registers();
-}
-
-// TODO(cd): Eliminate this file altogether when everything is in ASM file
diff --git a/vp9/common/arm/neon/vp9_short_idct32x32_add_neon.asm b/vp9/common/arm/neon/vp9_short_idct32x32_add_neon.asm
index 3a0ff60..b5a284b 100644
--- a/vp9/common/arm/neon/vp9_short_idct32x32_add_neon.asm
+++ b/vp9/common/arm/neon/vp9_short_idct32x32_add_neon.asm
@@ -43,8 +43,7 @@
cospi_31_64 EQU 804
- EXPORT |idct32_transpose_and_transform|
- EXPORT |idct32_combine_add|
+ EXPORT |vp9_short_idct32x32_add_neon|
ARM
REQUIRE8
PRESERVE8
@@ -100,6 +99,142 @@
vst1.16 {$reg2}, [r1]
MEND
; --------------------------------------------------------------------------
+ ; Combine-add results with current destination content
+ ; q6-q9 contain the results (out[j * 32 + 0-31])
+ MACRO
+ STORE_COMBINE_CENTER_RESULTS
+ ; load dest[j * dest_stride + 0-31]
+ vld1.s16 {d8}, [r10], r2
+ vld1.s16 {d11}, [r9], r11
+ vld1.s16 {d9}, [r10]
+ vld1.s16 {d10}, [r9]
+ ; ROUND_POWER_OF_TWO
+ vrshr.s16 q7, q7, #6
+ vrshr.s16 q8, q8, #6
+ vrshr.s16 q9, q9, #6
+ vrshr.s16 q6, q6, #6
+ ; add to dest[j * dest_stride + 0-31]
+ vaddw.u8 q7, q7, d9
+ vaddw.u8 q8, q8, d10
+ vaddw.u8 q9, q9, d11
+ vaddw.u8 q6, q6, d8
+ ; clip pixel
+ vqmovun.s16 d9, q7
+ vqmovun.s16 d10, q8
+ vqmovun.s16 d11, q9
+ vqmovun.s16 d8, q6
+ ; store back into dest[j * dest_stride + 0-31]
+ vst1.16 {d9}, [r10], r11
+ vst1.16 {d10}, [r9], r2
+ vst1.16 {d8}, [r10]
+ vst1.16 {d11}, [r9]
+ ; update pointers (by dest_stride * 2)
+ sub r9, r9, r2, lsl #1
+ add r10, r10, r2, lsl #1
+ MEND
+ ; --------------------------------------------------------------------------
+ ; Combine-add results with current destination content
+ ; q6-q9 contain the results (out[j * 32 + 0-31])
+ MACRO
+ STORE_COMBINE_CENTER_RESULTS_LAST
+ ; load dest[j * dest_stride + 0-31]
+ vld1.s16 {d8}, [r10], r2
+ vld1.s16 {d11}, [r9], r11
+ vld1.s16 {d9}, [r10]
+ vld1.s16 {d10}, [r9]
+ ; ROUND_POWER_OF_TWO
+ vrshr.s16 q7, q7, #6
+ vrshr.s16 q8, q8, #6
+ vrshr.s16 q9, q9, #6
+ vrshr.s16 q6, q6, #6
+ ; add to dest[j * dest_stride + 0-31]
+ vaddw.u8 q7, q7, d9
+ vaddw.u8 q8, q8, d10
+ vaddw.u8 q9, q9, d11
+ vaddw.u8 q6, q6, d8
+ ; clip pixel
+ vqmovun.s16 d9, q7
+ vqmovun.s16 d10, q8
+ vqmovun.s16 d11, q9
+ vqmovun.s16 d8, q6
+ ; store back into dest[j * dest_stride + 0-31]
+ vst1.16 {d9}, [r10], r11
+ vst1.16 {d10}, [r9], r2
+ vst1.16 {d8}, [r10]!
+ vst1.16 {d11}, [r9]!
+ ; update pointers (by dest_stride * 2)
+ sub r9, r9, r2, lsl #1
+ add r10, r10, r2, lsl #1
+ MEND
+ ; --------------------------------------------------------------------------
+ ; Combine-add results with current destination content
+ ; q4-q7 contain the results (out[j * 32 + 0-31])
+ MACRO
+ STORE_COMBINE_EXTREME_RESULTS
+ ; load dest[j * dest_stride + 0-31]
+ vld1.s16 {d4}, [r7], r2
+ vld1.s16 {d7}, [r6], r11
+ vld1.s16 {d5}, [r7]
+ vld1.s16 {d6}, [r6]
+ ; ROUND_POWER_OF_TWO
+ vrshr.s16 q5, q5, #6
+ vrshr.s16 q6, q6, #6
+ vrshr.s16 q7, q7, #6
+ vrshr.s16 q4, q4, #6
+ ; add to dest[j * dest_stride + 0-31]
+ vaddw.u8 q5, q5, d5
+ vaddw.u8 q6, q6, d6
+ vaddw.u8 q7, q7, d7
+ vaddw.u8 q4, q4, d4
+ ; clip pixel
+ vqmovun.s16 d5, q5
+ vqmovun.s16 d6, q6
+ vqmovun.s16 d7, q7
+ vqmovun.s16 d4, q4
+ ; store back into dest[j * dest_stride + 0-31]
+ vst1.16 {d5}, [r7], r11
+ vst1.16 {d6}, [r6], r2
+ vst1.16 {d7}, [r6]
+ vst1.16 {d4}, [r7]
+ ; update pointers (by dest_stride * 2)
+ sub r6, r6, r2, lsl #1
+ add r7, r7, r2, lsl #1
+ MEND
+ ; --------------------------------------------------------------------------
+ ; Combine-add results with current destination content
+ ; q4-q7 contain the results (out[j * 32 + 0-31])
+ MACRO
+ STORE_COMBINE_EXTREME_RESULTS_LAST
+ ; load dest[j * dest_stride + 0-31]
+ vld1.s16 {d4}, [r7], r2
+ vld1.s16 {d7}, [r6], r11
+ vld1.s16 {d5}, [r7]
+ vld1.s16 {d6}, [r6]
+ ; ROUND_POWER_OF_TWO
+ vrshr.s16 q5, q5, #6
+ vrshr.s16 q6, q6, #6
+ vrshr.s16 q7, q7, #6
+ vrshr.s16 q4, q4, #6
+ ; add to dest[j * dest_stride + 0-31]
+ vaddw.u8 q5, q5, d5
+ vaddw.u8 q6, q6, d6
+ vaddw.u8 q7, q7, d7
+ vaddw.u8 q4, q4, d4
+ ; clip pixel
+ vqmovun.s16 d5, q5
+ vqmovun.s16 d6, q6
+ vqmovun.s16 d7, q7
+ vqmovun.s16 d4, q4
+ ; store back into dest[j * dest_stride + 0-31]
+ vst1.16 {d5}, [r7], r11
+ vst1.16 {d6}, [r6], r2
+ vst1.16 {d7}, [r6]!
+ vst1.16 {d4}, [r7]!
+ ; update pointers (by dest_stride * 2)
+ sub r6, r6, r2, lsl #1
+ add r7, r7, r2, lsl #1
+ MEND
+ ; --------------------------------------------------------------------------
; Touches q8-q12, q15 (q13-q14 are preserved)
; valid output registers are anything but q8-q11
MACRO
@@ -110,12 +245,12 @@
; additions/substractions before the multiplies.
; generate the constants
; generate scalar constants
- mov r3, #$first_constant & 0xFF00
- add r3, #$first_constant & 0x00FF
+ mov r8, #$first_constant & 0xFF00
mov r12, #$second_constant & 0xFF00
+ add r8, #$first_constant & 0x00FF
add r12, #$second_constant & 0x00FF
; generate vector constants
- vdup.16 d30, r3
+ vdup.16 d30, r8
vdup.16 d31, r12
; (used) two for inputs (regA-regD), one for constants (q15)
; do some multiplications (ordered for maximum latency hiding)
@@ -153,15 +288,22 @@
MEND
; --------------------------------------------------------------------------
-;void idct32_transpose_and_transform(int16_t *transpose_buffer, int16_t *output, int16_t *input);
+;void vp9_short_idct32x32_add_neon(int16_t *input, uint8_t *dest, int dest_stride);
;
-; r0 int16_t *transpose_buffer
-; r1 int16_t *output
-; r2 int16_t *input)
-; TODO(cd): have more logical parameter ordering but this issue will disappear
-; when functions are combined.
+; r0 int16_t *input,
+; r1 uint8_t *dest,
+; r2 int dest_stride)
+; loop counters
+; r4 bands loop counter
+; r5 pass loop counter
+; r8 transpose loop counter
+; combine-add pointers
+; r6 dest + 31 * dest_stride, descending (30, 29, 28, ...)
+; r7 dest + 0 * dest_stride, ascending (1, 2, 3, ...)
+; r9 dest + 15 * dest_stride, descending (14, 13, 12, ...)
+; r10 dest + 16 * dest_stride, ascending (17, 18, 19, ...)
-|idct32_transpose_and_transform| PROC
+|vp9_short_idct32x32_add_neon| PROC
; This function does one pass of idct32x32 transform.
;
; This is done by transposing the input and then doing a 1d transform on
@@ -171,43 +313,73 @@
; The 1d transform is done by looping over bands of eight columns (the
; idct32_bands loop). For each band, the transform input transposition
; is done on demand, one band of four 8x8 matrices at a time. The four
- ; matrices are trsnposed by pairs (the idct32_transpose_pair loop).
- push {r4}
- mov r4, #0 ; initialize bands loop counter
+ ; matrices are transposed by pairs (the idct32_transpose_pair loop).
+ push {r4-r11}
+ vpush {d8-d15}
+ ; stack operation
+ ; internal buffer used to transpose 8 lines into before transforming them
+ ; int16_t transpose_buffer[32 * 8];
+ ; at sp + [4096, 4607]
+ ; results of the first pass (transpose and transform rows)
+ ; int16_t pass1[32 * 32];
+ ; at sp + [0, 2047]
+ ; results of the second pass (transpose and transform columns)
+ ; int16_t pass2[32 * 32];
+ ; at sp + [2048, 4095]
+ sub sp, sp, #512+2048+2048
+
+ ; r6 = dest + 31 * dest_stride
+ ; r7 = dest + 0 * dest_stride
+ ; r9 = dest + 15 * dest_stride
+ ; r10 = dest + 16 * dest_stride
+ rsb r6, r2, r2, lsl #5
+ rsb r9, r2, r2, lsl #4
+ add r10, r1, r2, lsl #4
+ mov r7, r1
+ add r6, r6, r1
+ add r9, r9, r1
+ ; r11 = -dest_stride
+ neg r11, r2
+ ; r3 = input
+ mov r3, r0
+ ; parameters for first pass
+ ; r0 = transpose_buffer[32 * 8]
+ add r0, sp, #4096
+ ; r1 = pass1[32 * 32]
+ mov r1, sp
+
+ mov r5, #0 ; initialize pass loop counter
+idct32_pass_loop
+ mov r4, #4 ; initialize bands loop counter
idct32_bands_loop
- ; TODO(cd) get rid of these push/pop by properly adjusting register
- ; content at end of loop
- push {r0}
- push {r1}
- push {r2}
- mov r3, #0 ; initialize transpose loop counter
+ mov r8, #2 ; initialize transpose loop counter
idct32_transpose_pair_loop
; Load two horizontally consecutive 8x8 16bit data matrices. The first one
; into q0-q7 and the second one into q8-q15. There is a stride of 64,
; adjusted to 32 because of the two post-increments.
- vld1.s16 {q8}, [r2]!
- vld1.s16 {q0}, [r2]!
- add r2, #32
- vld1.s16 {q9}, [r2]!
- vld1.s16 {q1}, [r2]!
- add r2, #32
- vld1.s16 {q10}, [r2]!
- vld1.s16 {q2}, [r2]!
- add r2, #32
- vld1.s16 {q11}, [r2]!
- vld1.s16 {q3}, [r2]!
- add r2, #32
- vld1.s16 {q12}, [r2]!
- vld1.s16 {q4}, [r2]!
- add r2, #32
- vld1.s16 {q13}, [r2]!
- vld1.s16 {q5}, [r2]!
- add r2, #32
- vld1.s16 {q14}, [r2]!
- vld1.s16 {q6}, [r2]!
- add r2, #32
- vld1.s16 {q15}, [r2]!
- vld1.s16 {q7}, [r2]!
+ vld1.s16 {q8}, [r3]!
+ vld1.s16 {q0}, [r3]!
+ add r3, #32
+ vld1.s16 {q9}, [r3]!
+ vld1.s16 {q1}, [r3]!
+ add r3, #32
+ vld1.s16 {q10}, [r3]!
+ vld1.s16 {q2}, [r3]!
+ add r3, #32
+ vld1.s16 {q11}, [r3]!
+ vld1.s16 {q3}, [r3]!
+ add r3, #32
+ vld1.s16 {q12}, [r3]!
+ vld1.s16 {q4}, [r3]!
+ add r3, #32
+ vld1.s16 {q13}, [r3]!
+ vld1.s16 {q5}, [r3]!
+ add r3, #32
+ vld1.s16 {q14}, [r3]!
+ vld1.s16 {q6}, [r3]!
+ add r3, #32
+ vld1.s16 {q15}, [r3]!
+ vld1.s16 {q7}, [r3]!
; Transpose the two 8x8 16bit data matrices.
vswp d17, d24
@@ -255,11 +427,13 @@
vst1.16 {q7}, [r0]!
; increment pointers by adjusted stride (not necessary for r0/out)
- sub r2, r2, #8*32*2-32-16*2
+ ; go back by 7*32 for the seven lines moved fully by read and add
+ ; go back by 32 for the eigth line only read
+ ; advance by 16*2 to go the next pair
+ sub r3, r3, #7*32*2 + 32 - 16*2
; transpose pair loop processing
- add r3, r3, #1
- cmp r3, #1
- ble idct32_transpose_pair_loop
+ subs r8, r8, #1
+ bne idct32_transpose_pair_loop
; restore r0/input to its original value
sub r0, r0, #32*8*2
@@ -815,21 +989,26 @@
vadd.s16 q9, q5, q0
vsub.s16 q6, q5, q0
vsub.s16 q7, q4, q1
- STORE_IN_OUTPUT 17, 17, 16, q7, q6
- STORE_IN_OUTPUT 16, 15, 14, q9, q8
+
+ cmp r5, #0
+ bgt idct32_bands_end_2nd_pass
+
+idct32_bands_end_1st_pass
+ STORE_IN_OUTPUT 17, 16, 17, q6, q7
+ STORE_IN_OUTPUT 17, 14, 15, q8, q9
; --------------------------------------------------------------------------
; part of final stage
;output[ 0 * 32] = step1b[0][i] + step1b[31][i];
;output[ 1 * 32] = step1b[1][i] + step1b[30][i];
;output[30 * 32] = step1b[1][i] - step1b[30][i];
;output[31 * 32] = step1b[0][i] - step1b[31][i];
- LOAD_FROM_OUTPUT 14, 30, 31, q0, q1
+ LOAD_FROM_OUTPUT 15, 30, 31, q0, q1
vadd.s16 q4, q2, q1
vadd.s16 q5, q3, q0
vsub.s16 q6, q3, q0
vsub.s16 q7, q2, q1
- STORE_IN_OUTPUT 31, 31, 30, q7, q6
- STORE_IN_OUTPUT 30, 0, 1, q4, q5
+ STORE_IN_OUTPUT 31, 30, 31, q6, q7
+ STORE_IN_OUTPUT 31, 0, 1, q4, q5
; --------------------------------------------------------------------------
; part of stage 7
;step1[2] = step1b[2][i] + step1b[13][i];
@@ -848,25 +1027,25 @@
;output[18 * 32] = step1b[13][i] - step1b[18][i];
;output[19 * 32] = step1b[12][i] - step1b[19][i];
LOAD_FROM_OUTPUT 13, 18, 19, q0, q1
- vadd.s16 q6, q4, q1
- vadd.s16 q7, q5, q0
- vsub.s16 q8, q5, q0
- vsub.s16 q9, q4, q1
- STORE_IN_OUTPUT 19, 19, 18, q9, q8
- STORE_IN_OUTPUT 18, 13, 12, q7, q6
+ vadd.s16 q8, q4, q1
+ vadd.s16 q9, q5, q0
+ vsub.s16 q6, q5, q0
+ vsub.s16 q7, q4, q1
+ STORE_IN_OUTPUT 19, 18, 19, q6, q7
+ STORE_IN_OUTPUT 19, 12, 13, q8, q9
; --------------------------------------------------------------------------
; part of final stage
;output[ 2 * 32] = step1b[2][i] + step1b[29][i];
;output[ 3 * 32] = step1b[3][i] + step1b[28][i];
;output[28 * 32] = step1b[3][i] - step1b[28][i];
;output[29 * 32] = step1b[2][i] - step1b[29][i];
- LOAD_FROM_OUTPUT 12, 28, 29, q0, q1
+ LOAD_FROM_OUTPUT 13, 28, 29, q0, q1
vadd.s16 q4, q2, q1
vadd.s16 q5, q3, q0
vsub.s16 q6, q3, q0
vsub.s16 q7, q2, q1
- STORE_IN_OUTPUT 29, 29, 28, q7, q6
- STORE_IN_OUTPUT 28, 2, 3, q4, q5
+ STORE_IN_OUTPUT 29, 28, 29, q6, q7
+ STORE_IN_OUTPUT 29, 2, 3, q4, q5
; --------------------------------------------------------------------------
; part of stage 7
;step1[4] = step1b[4][i] + step1b[11][i];
@@ -885,25 +1064,25 @@
;output[20 * 32] = step1b[11][i] - step1b[20][i];
;output[21 * 32] = step1b[10][i] - step1b[21][i];
LOAD_FROM_OUTPUT 11, 20, 21, q0, q1
- vadd.s16 q6, q4, q1
- vadd.s16 q7, q5, q0
- vsub.s16 q8, q5, q0
- vsub.s16 q9, q4, q1
- STORE_IN_OUTPUT 21, 21, 20, q9, q8
- STORE_IN_OUTPUT 20, 11, 10, q7, q6
+ vadd.s16 q8, q4, q1
+ vadd.s16 q9, q5, q0
+ vsub.s16 q6, q5, q0
+ vsub.s16 q7, q4, q1
+ STORE_IN_OUTPUT 21, 20, 21, q6, q7
+ STORE_IN_OUTPUT 21, 10, 11, q8, q9
; --------------------------------------------------------------------------
; part of final stage
;output[ 4 * 32] = step1b[4][i] + step1b[27][i];
;output[ 5 * 32] = step1b[5][i] + step1b[26][i];
;output[26 * 32] = step1b[5][i] - step1b[26][i];
;output[27 * 32] = step1b[4][i] - step1b[27][i];
- LOAD_FROM_OUTPUT 10, 26, 27, q0, q1
+ LOAD_FROM_OUTPUT 11, 26, 27, q0, q1
vadd.s16 q4, q2, q1
vadd.s16 q5, q3, q0
vsub.s16 q6, q3, q0
vsub.s16 q7, q2, q1
- STORE_IN_OUTPUT 27, 27, 26, q7, q6
- STORE_IN_OUTPUT 26, 4, 5, q4, q5
+ STORE_IN_OUTPUT 27, 26, 27, q6, q7
+ STORE_IN_OUTPUT 27, 4, 5, q4, q5
; --------------------------------------------------------------------------
; part of stage 7
;step1[6] = step1b[6][i] + step1b[9][i];
@@ -922,92 +1101,199 @@
;output[22 * 32] = step1b[9][i] - step1b[22][i];
;output[23 * 32] = step1b[8][i] - step1b[23][i];
LOAD_FROM_OUTPUT 9, 22, 23, q0, q1
- vadd.s16 q6, q4, q1
- vadd.s16 q7, q5, q0
- vsub.s16 q8, q5, q0
- vsub.s16 q9, q4, q1
- STORE_IN_OUTPUT 23, 23, 22, q9, q8
- STORE_IN_OUTPUT 22, 9, 8, q7, q6
+ vadd.s16 q8, q4, q1
+ vadd.s16 q9, q5, q0
+ vsub.s16 q6, q5, q0
+ vsub.s16 q7, q4, q1
+ STORE_IN_OUTPUT 23, 22, 23, q6, q7
+ STORE_IN_OUTPUT 23, 8, 9, q8, q9
; --------------------------------------------------------------------------
; part of final stage
;output[ 6 * 32] = step1b[6][i] + step1b[25][i];
;output[ 7 * 32] = step1b[7][i] + step1b[24][i];
;output[24 * 32] = step1b[7][i] - step1b[24][i];
;output[25 * 32] = step1b[6][i] - step1b[25][i];
- LOAD_FROM_OUTPUT 8, 24, 25, q0, q1
+ LOAD_FROM_OUTPUT 9, 24, 25, q0, q1
vadd.s16 q4, q2, q1
vadd.s16 q5, q3, q0
vsub.s16 q6, q3, q0
vsub.s16 q7, q2, q1
- STORE_IN_OUTPUT 25, 25, 24, q7, q6
- STORE_IN_OUTPUT 24, 6, 7, q4, q5
- ; --------------------------------------------------------------------------
+ STORE_IN_OUTPUT 25, 24, 25, q6, q7
+ STORE_IN_OUTPUT 25, 6, 7, q4, q5
- ; TODO(cd) get rid of these push/pop by properly adjusting register
- ; content at end of loop
- pop {r2}
- pop {r1}
- pop {r0}
- add r1, r1, #8*2
- add r2, r2, #8*32*2
+ ; restore r0 by removing the last offset from the last
+ ; operation (LOAD_FROM_TRANSPOSED 16, 8, 24) => 24*8*2
+ sub r0, r0, #24*8*2
+ ; restore r1 by removing the last offset from the last
+ ; operation (STORE_IN_OUTPUT 24, 6, 7) => 7*32*2
+ ; advance by 8 columns => 8*2
+ sub r1, r1, #7*32*2 - 8*2
+ ; advance by 8 lines (8*32*2)
+ ; go back by the two pairs from the loop (32*2)
+ add r3, r3, #8*32*2 - 32*2
; bands loop processing
- add r4, r4, #1
- cmp r4, #3
- ble idct32_bands_loop
+ subs r4, r4, #1
+ bne idct32_bands_loop
- pop {r4}
+ ; parameters for second pass
+ ; the input of pass2 is the result of pass1. we have to remove the offset
+ ; of 32 columns induced by the above idct32_bands_loop
+ sub r3, r1, #32*2
+ ; r1 = pass2[32 * 32]
+ add r1, sp, #2048
+
+ ; pass loop processing
+ add r5, r5, #1
+ B idct32_pass_loop
+
+idct32_bands_end_2nd_pass
+ STORE_COMBINE_CENTER_RESULTS
+ ; --------------------------------------------------------------------------
+ ; part of final stage
+ ;output[ 0 * 32] = step1b[0][i] + step1b[31][i];
+ ;output[ 1 * 32] = step1b[1][i] + step1b[30][i];
+ ;output[30 * 32] = step1b[1][i] - step1b[30][i];
+ ;output[31 * 32] = step1b[0][i] - step1b[31][i];
+ LOAD_FROM_OUTPUT 17, 30, 31, q0, q1
+ vadd.s16 q4, q2, q1
+ vadd.s16 q5, q3, q0
+ vsub.s16 q6, q3, q0
+ vsub.s16 q7, q2, q1
+ STORE_COMBINE_EXTREME_RESULTS
+ ; --------------------------------------------------------------------------
+ ; part of stage 7
+ ;step1[2] = step1b[2][i] + step1b[13][i];
+ ;step1[3] = step1b[3][i] + step1b[12][i];
+ ;step1[12] = step1b[3][i] - step1b[12][i];
+ ;step1[13] = step1b[2][i] - step1b[13][i];
+ LOAD_FROM_OUTPUT 31, 12, 13, q0, q1
+ vadd.s16 q2, q10, q1
+ vadd.s16 q3, q11, q0
+ vsub.s16 q4, q11, q0
+ vsub.s16 q5, q10, q1
+ ; --------------------------------------------------------------------------
+ ; part of final stage
+ ;output[12 * 32] = step1b[12][i] + step1b[19][i];
+ ;output[13 * 32] = step1b[13][i] + step1b[18][i];
+ ;output[18 * 32] = step1b[13][i] - step1b[18][i];
+ ;output[19 * 32] = step1b[12][i] - step1b[19][i];
+ LOAD_FROM_OUTPUT 13, 18, 19, q0, q1
+ vadd.s16 q8, q4, q1
+ vadd.s16 q9, q5, q0
+ vsub.s16 q6, q5, q0
+ vsub.s16 q7, q4, q1
+ STORE_COMBINE_CENTER_RESULTS
+ ; --------------------------------------------------------------------------
+ ; part of final stage
+ ;output[ 2 * 32] = step1b[2][i] + step1b[29][i];
+ ;output[ 3 * 32] = step1b[3][i] + step1b[28][i];
+ ;output[28 * 32] = step1b[3][i] - step1b[28][i];
+ ;output[29 * 32] = step1b[2][i] - step1b[29][i];
+ LOAD_FROM_OUTPUT 19, 28, 29, q0, q1
+ vadd.s16 q4, q2, q1
+ vadd.s16 q5, q3, q0
+ vsub.s16 q6, q3, q0
+ vsub.s16 q7, q2, q1
+ STORE_COMBINE_EXTREME_RESULTS
+ ; --------------------------------------------------------------------------
+ ; part of stage 7
+ ;step1[4] = step1b[4][i] + step1b[11][i];
+ ;step1[5] = step1b[5][i] + step1b[10][i];
+ ;step1[10] = step1b[5][i] - step1b[10][i];
+ ;step1[11] = step1b[4][i] - step1b[11][i];
+ LOAD_FROM_OUTPUT 29, 10, 11, q0, q1
+ vadd.s16 q2, q12, q1
+ vadd.s16 q3, q13, q0
+ vsub.s16 q4, q13, q0
+ vsub.s16 q5, q12, q1
+ ; --------------------------------------------------------------------------
+ ; part of final stage
+ ;output[10 * 32] = step1b[10][i] + step1b[21][i];
+ ;output[11 * 32] = step1b[11][i] + step1b[20][i];
+ ;output[20 * 32] = step1b[11][i] - step1b[20][i];
+ ;output[21 * 32] = step1b[10][i] - step1b[21][i];
+ LOAD_FROM_OUTPUT 11, 20, 21, q0, q1
+ vadd.s16 q8, q4, q1
+ vadd.s16 q9, q5, q0
+ vsub.s16 q6, q5, q0
+ vsub.s16 q7, q4, q1
+ STORE_COMBINE_CENTER_RESULTS
+ ; --------------------------------------------------------------------------
+ ; part of final stage
+ ;output[ 4 * 32] = step1b[4][i] + step1b[27][i];
+ ;output[ 5 * 32] = step1b[5][i] + step1b[26][i];
+ ;output[26 * 32] = step1b[5][i] - step1b[26][i];
+ ;output[27 * 32] = step1b[4][i] - step1b[27][i];
+ LOAD_FROM_OUTPUT 21, 26, 27, q0, q1
+ vadd.s16 q4, q2, q1
+ vadd.s16 q5, q3, q0
+ vsub.s16 q6, q3, q0
+ vsub.s16 q7, q2, q1
+ STORE_COMBINE_EXTREME_RESULTS
+ ; --------------------------------------------------------------------------
+ ; part of stage 7
+ ;step1[6] = step1b[6][i] + step1b[9][i];
+ ;step1[7] = step1b[7][i] + step1b[8][i];
+ ;step1[8] = step1b[7][i] - step1b[8][i];
+ ;step1[9] = step1b[6][i] - step1b[9][i];
+ LOAD_FROM_OUTPUT 27, 8, 9, q0, q1
+ vadd.s16 q2, q14, q1
+ vadd.s16 q3, q15, q0
+ vsub.s16 q4, q15, q0
+ vsub.s16 q5, q14, q1
+ ; --------------------------------------------------------------------------
+ ; part of final stage
+ ;output[ 8 * 32] = step1b[8][i] + step1b[23][i];
+ ;output[ 9 * 32] = step1b[9][i] + step1b[22][i];
+ ;output[22 * 32] = step1b[9][i] - step1b[22][i];
+ ;output[23 * 32] = step1b[8][i] - step1b[23][i];
+ LOAD_FROM_OUTPUT 9, 22, 23, q0, q1
+ vadd.s16 q8, q4, q1
+ vadd.s16 q9, q5, q0
+ vsub.s16 q6, q5, q0
+ vsub.s16 q7, q4, q1
+ STORE_COMBINE_CENTER_RESULTS_LAST
+ ; --------------------------------------------------------------------------
+ ; part of final stage
+ ;output[ 6 * 32] = step1b[6][i] + step1b[25][i];
+ ;output[ 7 * 32] = step1b[7][i] + step1b[24][i];
+ ;output[24 * 32] = step1b[7][i] - step1b[24][i];
+ ;output[25 * 32] = step1b[6][i] - step1b[25][i];
+ LOAD_FROM_OUTPUT 23, 24, 25, q0, q1
+ vadd.s16 q4, q2, q1
+ vadd.s16 q5, q3, q0
+ vsub.s16 q6, q3, q0
+ vsub.s16 q7, q2, q1
+ STORE_COMBINE_EXTREME_RESULTS_LAST
+ ; --------------------------------------------------------------------------
+ ; restore pointers to their initial indices for next band pass by
+ ; removing/adding dest_stride * 8. The actual increment by eight
+ ; is taken care of within the _LAST macros.
+ add r6, r6, r2, lsl #3
+ add r9, r9, r2, lsl #3
+ sub r7, r7, r2, lsl #3
+ sub r10, r10, r2, lsl #3
+
+ ; restore r0 by removing the last offset from the last
+ ; operation (LOAD_FROM_TRANSPOSED 16, 8, 24) => 24*8*2
+ sub r0, r0, #24*8*2
+ ; restore r1 by removing the last offset from the last
+ ; operation (LOAD_FROM_OUTPUT 23, 24, 25) => 25*32*2
+ ; advance by 8 columns => 8*2
+ sub r1, r1, #25*32*2 - 8*2
+ ; advance by 8 lines (8*32*2)
+ ; go back by the two pairs from the loop (32*2)
+ add r3, r3, #8*32*2 - 32*2
+
+ ; bands loop processing
+ subs r4, r4, #1
+ bne idct32_bands_loop
+
+ ; stack operation
+ add sp, sp, #512+2048+2048
+ vpop {d8-d15}
+ pop {r4-r11}
bx lr
- ENDP ; |idct32_transpose_and_transform|
-
-;void idct32_combine_add(uint8_t *dest, int16_t *out, int dest_stride);
-;
-; r0 uint8_t *dest
-; r1 int16_t *out
-; r2 int dest_stride)
-
-|idct32_combine_add| PROC
-
- mov r12, r0 ; dest pointer used for stores
- sub r2, r2, #32 ; adjust the stride (remove the post-increments)
- mov r3, #0 ; initialize loop counter
-
-idct32_combine_add_loop
- ; load out[j * 32 + 0-31]
- vld1.s16 {q12}, [r1]!
- vld1.s16 {q13}, [r1]!
- vld1.s16 {q14}, [r1]!
- vld1.s16 {q15}, [r1]!
- ; load dest[j * dest_stride + 0-31]
- vld1.s16 {q6}, [r0]!
- vld1.s16 {q7}, [r0]!
- ; ROUND_POWER_OF_TWO
- vrshr.s16 q12, q12, #6
- vrshr.s16 q13, q13, #6
- vrshr.s16 q14, q14, #6
- vrshr.s16 q15, q15, #6
- ; add to dest[j * dest_stride + 0-31]
- vaddw.u8 q12, q12, d12
- vaddw.u8 q13, q13, d13
- vaddw.u8 q14, q14, d14
- vaddw.u8 q15, q15, d15
- ; clip pixel
- vqmovun.s16 d12, q12
- vqmovun.s16 d13, q13
- vqmovun.s16 d14, q14
- vqmovun.s16 d15, q15
- ; store back into dest[j * dest_stride + 0-31]
- vst1.16 {q6}, [r12]!
- vst1.16 {q7}, [r12]!
- ; increment pointers by adjusted stride (not necessary for r1/out)
- add r0, r0, r2
- add r12, r12, r2
- ; loop processing
- add r3, r3, #1
- cmp r3, #31
- ble idct32_combine_add_loop
-
- bx lr
- ENDP ; |idct32_transpose|
-
+ ENDP ; |vp9_short_idct32x32_add_neon|
END
diff --git a/vp9/common/vp9_rtcd_defs.sh b/vp9/common/vp9_rtcd_defs.sh
index 52bcbcf..c4d7134 100644
--- a/vp9/common/vp9_rtcd_defs.sh
+++ b/vp9/common/vp9_rtcd_defs.sh
@@ -53,7 +53,7 @@
specialize vp9_d45_predictor_4x4 $ssse3_x86inc
prototype void vp9_d63_predictor_4x4 "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left"
-specialize vp9_d63_predictor_4x4
+specialize vp9_d63_predictor_4x4 $ssse3_x86inc
prototype void vp9_h_predictor_4x4 "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left"
specialize vp9_h_predictor_4x4 $ssse3_x86inc
@@ -92,7 +92,7 @@
specialize vp9_d45_predictor_8x8 $ssse3_x86inc
prototype void vp9_d63_predictor_8x8 "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left"
-specialize vp9_d63_predictor_8x8
+specialize vp9_d63_predictor_8x8 $ssse3_x86inc
prototype void vp9_h_predictor_8x8 "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left"
specialize vp9_h_predictor_8x8 $ssse3_x86inc
@@ -131,7 +131,7 @@
specialize vp9_d45_predictor_16x16 $ssse3_x86inc
prototype void vp9_d63_predictor_16x16 "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left"
-specialize vp9_d63_predictor_16x16
+specialize vp9_d63_predictor_16x16 $ssse3_x86inc
prototype void vp9_h_predictor_16x16 "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left"
specialize vp9_h_predictor_16x16 $ssse3_x86inc
@@ -170,7 +170,7 @@
specialize vp9_d45_predictor_32x32 $ssse3_x86inc
prototype void vp9_d63_predictor_32x32 "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left"
-specialize vp9_d63_predictor_32x32
+specialize vp9_d63_predictor_32x32 $ssse3_x86inc
prototype void vp9_h_predictor_32x32 "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left"
specialize vp9_h_predictor_32x32 $ssse3 x86inc
diff --git a/vp9/common/x86/vp9_intrapred_ssse3.asm b/vp9/common/x86/vp9_intrapred_ssse3.asm
index 67c8ab0..c51d011 100644
--- a/vp9/common/x86/vp9_intrapred_ssse3.asm
+++ b/vp9/common/x86/vp9_intrapred_ssse3.asm
@@ -17,8 +17,8 @@
pb_7m1: times 8 db 7, -1
pb_15: times 16 db 15
-sh_b01234577: db 0, 1, 2, 3, 4, 5, 7, 7
-sh_b12345677: db 1, 2, 3, 4, 5, 6, 7, 7
+sh_b01234577: db 0, 1, 2, 3, 4, 5, 7, 7, 0, 0, 0, 0, 0, 0, 0, 0
+sh_b12345677: db 1, 2, 3, 4, 5, 6, 7, 7, 0, 0, 0, 0, 0, 0, 0, 0
sh_b23456777: db 2, 3, 4, 5, 6, 7, 7, 7, 0, 0, 0, 0, 0, 0, 0, 0
sh_b0123456777777777: db 0, 1, 2, 3, 4, 5, 6, 7, 7, 7, 7, 7, 7, 7, 7, 7
sh_b1234567777777777: db 1, 2, 3, 4, 5, 6, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7
@@ -305,3 +305,153 @@
RESTORE_GOT
RET
+
+; ------------------------------------------
+; input: x, y, z, result
+;
+; trick from pascal
+; (x+2y+z+2)>>2 can be calculated as:
+; result = avg(x,z)
+; result -= xor(x,z) & 1
+; result = avg(result,y)
+; ------------------------------------------
+%macro X_PLUS_2Y_PLUS_Z_PLUS_2_RSH_2 4
+ pavgb %4, %1, %3
+ pxor %3, %1
+ pand %3, [GLOBAL(pb_1)]
+ psubb %4, %3
+ pavgb %4, %2
+%endmacro
+
+INIT_XMM ssse3
+cglobal d63_predictor_4x4, 3, 4, 5, dst, stride, above, goffset
+ GET_GOT goffsetq
+
+ movq m3, [aboveq]
+ pshufb m1, m3, [GLOBAL(sh_b23456777)]
+ pshufb m2, m3, [GLOBAL(sh_b12345677)]
+
+ X_PLUS_2Y_PLUS_Z_PLUS_2_RSH_2 m3, m2, m1, m4
+ pavgb m3, m2
+
+ ; store 4 lines
+ movd [dstq ], m3
+ movd [dstq+strideq], m4
+ lea dstq, [dstq+strideq*2]
+ psrldq m3, 1
+ psrldq m4, 1
+ movd [dstq ], m3
+ movd [dstq+strideq], m4
+ RESTORE_GOT
+ RET
+
+INIT_XMM ssse3
+cglobal d63_predictor_8x8, 3, 4, 5, dst, stride, above, goffset
+ GET_GOT goffsetq
+
+ movq m3, [aboveq]
+ DEFINE_ARGS dst, stride, stride3
+ lea stride3q, [strideq*3]
+ pshufb m1, m3, [GLOBAL(sh_b2345677777777777)]
+ pshufb m0, m3, [GLOBAL(sh_b0123456777777777)]
+ pshufb m2, m3, [GLOBAL(sh_b1234567777777777)]
+ pshufb m3, [GLOBAL(sh_b0123456777777777)]
+
+ X_PLUS_2Y_PLUS_Z_PLUS_2_RSH_2 m0, m2, m1, m4
+ pavgb m3, m2
+
+ ; store 4 lines
+ movq [dstq ], m3
+ movq [dstq+strideq], m4
+ psrldq m3, 1
+ psrldq m4, 1
+ movq [dstq+strideq*2], m3
+ movq [dstq+stride3q ], m4
+ lea dstq, [dstq+strideq*4]
+ psrldq m3, 1
+ psrldq m4, 1
+
+ ; store 4 lines
+ movq [dstq ], m3
+ movq [dstq+strideq], m4
+ psrldq m3, 1
+ psrldq m4, 1
+ movq [dstq+strideq*2], m3
+ movq [dstq+stride3q ], m4
+ RESTORE_GOT
+ RET
+
+INIT_XMM ssse3
+cglobal d63_predictor_16x16, 3, 5, 5, dst, stride, above, line, goffset
+ GET_GOT goffsetq
+
+ mova m0, [aboveq]
+ DEFINE_ARGS dst, stride, stride3, line
+ lea stride3q, [strideq*3]
+ mova m1, [GLOBAL(sh_b123456789abcdeff)]
+ pshufb m2, m0, [GLOBAL(sh_b23456789abcdefff)]
+ pshufb m3, m0, m1
+
+ X_PLUS_2Y_PLUS_Z_PLUS_2_RSH_2 m0, m3, m2, m4
+ pavgb m0, m3
+
+ mov lined, 4
+.loop:
+ mova [dstq ], m0
+ mova [dstq+strideq ], m4
+ pshufb m0, m1
+ pshufb m4, m1
+ mova [dstq+strideq*2], m0
+ mova [dstq+stride3q ], m4
+ pshufb m0, m1
+ pshufb m4, m1
+ lea dstq, [dstq+strideq*4]
+ dec lined
+ jnz .loop
+ RESTORE_GOT
+ REP_RET
+
+INIT_XMM ssse3
+cglobal d63_predictor_32x32, 3, 5, 8, dst, stride, above, line, goffset
+ GET_GOT goffsetq
+
+ mova m0, [aboveq]
+ mova m7, [aboveq+16]
+ DEFINE_ARGS dst, stride, stride3, line
+ mova m1, [GLOBAL(sh_b123456789abcdeff)]
+ lea stride3q, [strideq*3]
+ pshufb m2, m7, [GLOBAL(sh_b23456789abcdefff)]
+ pshufb m3, m7, m1
+
+ X_PLUS_2Y_PLUS_Z_PLUS_2_RSH_2 m7, m3, m2, m4
+ palignr m6, m7, m0, 1
+ palignr m5, m7, m0, 2
+ pavgb m7, m3
+
+ X_PLUS_2Y_PLUS_Z_PLUS_2_RSH_2 m0, m6, m5, m2
+ pavgb m0, m6
+
+ mov lined, 8
+.loop:
+ mova [dstq ], m0
+ mova [dstq +16], m7
+ mova [dstq+strideq ], m2
+ mova [dstq+strideq +16], m4
+ palignr m3, m7, m0, 1
+ palignr m5, m4, m2, 1
+ pshufb m7, m1
+ pshufb m4, m1
+
+ mova [dstq+strideq*2 ], m3
+ mova [dstq+strideq*2+16], m7
+ mova [dstq+stride3q ], m5
+ mova [dstq+stride3q +16], m4
+ palignr m0, m7, m3, 1
+ palignr m2, m4, m5, 1
+ pshufb m7, m1
+ pshufb m4, m1
+ lea dstq, [dstq+strideq*4]
+ dec lined
+ jnz .loop
+ RESTORE_GOT
+ REP_RET
diff --git a/vp9/decoder/vp9_decodemv.c b/vp9/decoder/vp9_decodemv.c
index dc12876..cef5ada 100644
--- a/vp9/decoder/vp9_decodemv.c
+++ b/vp9/decoder/vp9_decodemv.c
@@ -75,28 +75,9 @@
return TX_4X4;
}
-static void set_segment_id(VP9_COMMON *cm, BLOCK_SIZE bsize,
- int mi_row, int mi_col, int segment_id) {
- const int mi_offset = mi_row * cm->mi_cols + mi_col;
- const int bw = 1 << mi_width_log2(bsize);
- const int bh = 1 << mi_height_log2(bsize);
- const int xmis = MIN(cm->mi_cols - mi_col, bw);
- const int ymis = MIN(cm->mi_rows - mi_row, bh);
- int x, y;
-
- assert(segment_id >= 0 && segment_id < MAX_SEGMENTS);
-
- for (y = 0; y < ymis; y++)
- for (x = 0; x < xmis; x++)
- cm->last_frame_seg_map[mi_offset + y * cm->mi_cols + x] = segment_id;
-}
-
static int read_intra_segment_id(VP9D_COMP *pbi, int mi_row, int mi_col,
vp9_reader *r) {
- MACROBLOCKD *const xd = &pbi->mb;
struct segmentation *const seg = &pbi->common.seg;
- const BLOCK_SIZE bsize = xd->this_mi->mbmi.sb_type;
- int segment_id;
if (!seg->enabled)
return 0; // Default for disabled segmentation
@@ -104,9 +85,7 @@
if (!seg->update_map)
return 0;
- segment_id = read_segment_id(r, seg);
- set_segment_id(&pbi->common, bsize, mi_row, mi_col, segment_id);
- return segment_id;
+ return read_segment_id(r, seg);
}
static int read_inter_segment_id(VP9D_COMP *pbi, int mi_row, int mi_col,
@@ -115,7 +94,7 @@
MACROBLOCKD *const xd = &pbi->mb;
struct segmentation *const seg = &cm->seg;
const BLOCK_SIZE bsize = xd->this_mi->mbmi.sb_type;
- int pred_segment_id, segment_id;
+ int pred_segment_id;;
if (!seg->enabled)
return 0; // Default for disabled segmentation
@@ -129,13 +108,10 @@
const vp9_prob pred_prob = vp9_get_pred_prob_seg_id(seg, xd);
const int pred_flag = vp9_read(r, pred_prob);
vp9_set_pred_flag_seg_id(xd, pred_flag);
- segment_id = pred_flag ? pred_segment_id
- : read_segment_id(r, seg);
+ return pred_flag ? pred_segment_id : read_segment_id(r, seg);
} else {
- segment_id = read_segment_id(r, seg);
+ return read_segment_id(r, seg);
}
- set_segment_id(cm, bsize, mi_row, mi_col, segment_id);
- return segment_id;
}
static uint8_t read_skip_coeff(VP9D_COMP *pbi, int segment_id, vp9_reader *r) {
diff --git a/vp9/decoder/vp9_decodframe.c b/vp9/decoder/vp9_decodframe.c
index 34ed0c7..2331bb5 100644
--- a/vp9/decoder/vp9_decodframe.c
+++ b/vp9/decoder/vp9_decodframe.c
@@ -935,6 +935,15 @@
}
}
+static void update_segmentation_map(VP9_COMMON *cm) {
+ int i, j;
+
+ for (i = 0; i < cm->mi_rows; ++i)
+ for (j = 0; j < cm->mi_cols; ++j)
+ cm->last_frame_seg_map[i * cm->mi_cols + j] =
+ cm->mi_grid_visible[i * cm->mode_info_stride + j]->mbmi.segment_id;
+}
+
int vp9_decode_frame(VP9D_COMP *pbi, const uint8_t **p_data_end) {
int i;
VP9_COMMON *const cm = &pbi->common;
@@ -1014,5 +1023,7 @@
if (cm->refresh_frame_context)
cm->frame_contexts[cm->frame_context_idx] = cm->fc;
+ update_segmentation_map(cm);
+
return 0;
}
diff --git a/vp9/vp9_common.mk b/vp9/vp9_common.mk
index 687fb48..6778474 100644
--- a/vp9/vp9_common.mk
+++ b/vp9/vp9_common.mk
@@ -92,7 +92,6 @@
VP9_COMMON_SRCS-$(HAVE_NEON) += common/arm/neon/vp9_convolve_neon.c
VP9_COMMON_SRCS-$(HAVE_NEON) += common/arm/neon/vp9_idct16x16_neon.c
-VP9_COMMON_SRCS-$(HAVE_NEON) += common/arm/neon/vp9_idct32x32_neon.c
VP9_COMMON_SRCS-$(HAVE_NEON) += common/arm/neon/vp9_convolve8_neon$(ASM)
VP9_COMMON_SRCS-$(HAVE_NEON) += common/arm/neon/vp9_convolve8_avg_neon$(ASM)
VP9_COMMON_SRCS-$(HAVE_NEON) += common/arm/neon/vp9_loopfilter_neon$(ASM)
